Deep learning segmentation, classification, and risk prediction of complex vascular lesions on intravascular ultrasound images

Abstract

BackgroundThis study aimed to access the performance of our deep learning-based model in the segmentation and classification of vascular lesions on IVUS images. MethodsA total of 5,089 IVUS frames derived from 100 patients with stable or unstable angina pectoris were retrospectively collected. Our deep learning diagnostic framework composed of two stages, image segmentation and lesion classification. Segmentation was performed using the dilated attention U-Net model. The standard classifier, ResNet18, was subsequently used to classify the lesions into six groups, plaques (fibrous, lipid, and calcified), dissections, hematomas, and thrombi. Segmentation performance was evaluated in terms of Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95), while classification was assessed based on sensitivity, specificity, area under the receiver operating characteristic curve, and F1 score. The predictive value of our proposed model for high-risk lesions based on geometric measurements was further assessed. ResultsThe segmentation performance of our model was deemed satisfactory, with average DSC values of 80.75%, 86.68%, and 79.21% demonstrated for dissections, hematomas, and thrombi, respectively. Good lesion classification performance was observed as well, with F1 scores of 94.89%, 95.91% and 96.42%, respectively. Our model further demonstrated the ability to stratify lesions at risk for dissection. ConclusionOur deep-learning diagnostic framework demonstrated accuracy in the identification, classification, and risk stratification of vascular lesions based on IVUS images. It is clinically conducive, easily adoptable, and enables the early diagnosis of complex lesions at risk of major adverse cardiovascular events.